app.py

import streamlit as st
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import difflib
import requests
import os
import json

FIREBASE_URL = os.getenv("FIREBASE_URL")


def fetch_from_firebase(model_id, data_type):
    response = requests.get(f"{FIREBASE_URL}/{data_type}/{model_id}.json")
    if response.status_code == 200:
        return response.json()
    return None


def save_to_firebase(model_id, data, data_type):
    response = requests.put(
        f"{FIREBASE_URL}/{data_type}/{model_id}.json", data=json.dumps(data)
    )
    return response.status_code == 200


def get_model_structure(model_id) -> list[str]:
    struct_lines = fetch_from_firebase(model_id, "model_structures")
    if struct_lines:
        return struct_lines
    model = AutoModelForCausalLM.from_pretrained(
        model_id, torch_dtype=torch.bfloat16, device_map="cpu", trust_remote_code=True
    )
    structure = {k: str(v.shape) for k, v in model.state_dict().items()}
    struct_lines = [f"{k}: {v}" for k, v in structure.items()]
    save_to_firebase(model_id, struct_lines, "model_structures")
    return struct_lines


def get_tokenizer_vocab_size(model_id) -> int:
    vocab_size = fetch_from_firebase(model_id, "tokenizer_vocab_sizes")
    if vocab_size:
        return vocab_size
    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
    vocab_size = tokenizer.vocab_size
    save_to_firebase(model_id, vocab_size, "tokenizer_vocab_sizes")
    return vocab_size


def compare_structures(struct1_lines: list[str], struct2_lines: list[str]):
    diff = difflib.ndiff(struct1_lines, struct2_lines)
    return diff


def display_diff(diff):
    left_lines = []
    right_lines = []
    diff_found = False

    for line in diff:
        if line.startswith("- "):
            left_lines.append(
                f'<span style="background-color: #ffdddd;">{line[2:]}</span>'
            )
            right_lines.append("")
            diff_found = True
        elif line.startswith("+ "):
            right_lines.append(
                f'<span style="background-color: #ddffdd;">{line[2:]}</span>'
            )
            left_lines.append("")
            diff_found = True
        elif line.startswith("  "):
            left_lines.append(line[2:])
            right_lines.append(line[2:])
        else:
            pass

    left_html = "<br>".join(left_lines)
    right_html = "<br>".join(right_lines)

    return left_html, right_html, diff_found


# Set Streamlit page configuration to wide mode
st.set_page_config(layout="wide")

# Apply custom CSS for wider layout
st.markdown(
    """
    <style>
    .reportview-container .main .block-container {
        max-width: 100%;
        padding-left: 10%;
        padding-right: 10%;
    }
    .stMarkdown {
        white-space: pre-wrap;
    }
    </style>
    """,
    unsafe_allow_html=True,
)

st.title("Model Structure Comparison Tool")
model_id1 = st.text_input("Enter the first HuggingFace Model ID")
model_id2 = st.text_input("Enter the second HuggingFace Model ID")

if st.button("Compare Models"):
    with st.spinner("Comparing models and loading tokenizers..."):
        if model_id1 and model_id2:
            # Get model structures
            struct1 = get_model_structure(model_id1)
            struct2 = get_model_structure(model_id2)

            # Compare model structures
            diff = compare_structures(struct1, struct2)
            left_html, right_html, diff_found = display_diff(diff)

            st.write("### Comparison Result")
            if not diff_found:
                st.success("The model structures are identical.")

            col1, col2 = st.columns(
                [1.5, 1.5]
            )  # Adjust the ratio to make columns wider

            with col1:
                st.write(f"### Model 1: {model_id1}")
                st.markdown(left_html, unsafe_allow_html=True)

            with col2:
                st.write(f"### Model 2: {model_id2}")
                st.markdown(right_html, unsafe_allow_html=True)

            # Tokenizer verification
            try:
                vocab_size1 = get_tokenizer_vocab_size(model_id1)
                vocab_size2 = get_tokenizer_vocab_size(model_id2)

                if vocab_size1 == vocab_size2:
                    st.success("The tokenizer vocab sizes are identical.")
                else:
                    st.warning("The tokenizer vocab sizes are different.")

                st.write(f"**{model_id1} Tokenizer Vocab Size**: {vocab_size1}")
                st.write(f"**{model_id2} Tokenizer Vocab Size**: {vocab_size2}")

            except Exception as e:
                st.error(f"Error loading tokenizers: {e}")
        else:
            st.error("Please enter both model IDs.")